STRC Research

Phase 7I — Boltz-2 v52 combined analysis (top3 + rest11)

4 min read

Phase 7I — Boltz-2 v52 combined analysis (top3 + rest11)

Closes the v52_top3 batch (3 jobs DELIVERED 2026-04-25, never analyzed) and unifies it with the previously-analyzed v52_rest11 (9 paired ligands, Phase 7H follow-on logged earlier this session). Same scaffold throughout: 2-amino-quinoline-3 (aq3), an anthranilate-NH-aryl carrier per Phase 7H classification.

Script: scripts/phase7i_v52_combined_analyze.py. Outputs: v52_combined_results.{json,csv}.

Headline

  • 12 paired ligands, 4 mut-prefer · 5 WT-prefer · 3 tie. Mean Δ = −0.007 ± 0.055 (n=12), one-sample t = −0.43, df=11. The blanket “aq3 carrier → mut-prefer” hypothesis is FALSIFIED for v5.2 chemistry. Carrier-only is not the discriminator.
  • Head-group is the discriminator, not carrier. Decomposition (Δ in ligand_iptm units, mean ± SD):
Headnmean Δsdcall
CONHOMe2+0.0660.001🟢
phosphonate2+0.0500.023🟢
CONHOH2−0.0080.012⚪
CONHSO2Me4−0.0430.037🔴
COOH2−0.0620.002🔴
  • Lead spotlight — adamantyl_CONHOH_-Cl (the v5.2 paper-claim lead): WT 0.525 ± 0.068, E1659A 0.508 ± 0.114, Δ = −0.017 → ⚪ tie. The lead does not fall in the 🟢 mut-prefer cluster on Boltz-2.
  • New top hit on adamantyl body: adamantyl_CONHOMe_-Cl Δ = +0.066, adamantyl_CONHOMe_-CN Δ = +0.065 — confirms CONHOMe-head adamantyl variant across both ortho substituents (-Cl and -CN both 🟢).

What this changes vs Phase 7H rest11 alone

  • v52_top3 closes: 3 new paired ligands (adamantyl_CONHOH_-Cl, adamantyl_CONHOMe_-Cl, 1-indanyl_COOH_-Me) integrated.
  • The CONHOMe class moves from “1 geometry, needs more variants” (rest11 log) → “2 geometries, both 🟢, mean Δ = +0.066, sd 0.001” — internally tight class signal.
  • The lead adamantyl_CONHOH_-Cl is now characterized: tie (Δ = −0.017). Phase 5k mech-4 proof is on the +4.53 kT/e pocket attraction (formal-anion electrostatic component, decomposed thermodynamic), which holds. Boltz-2 measures full binding likelihood (entropy + dynamics + scoring) — the lead’s hydroxamic-acid head doesn’t translate the pocket attraction into a Boltz-2 mut-preference signal.
  • Phase 7H rule refinement: not “anthranilate-NH-aryl carrier → mut-prefer,” but “anthranilate-NH-aryl carrier × {CONHOMe, phosphonate} head → mut-prefer.” CONHOH is necessary at the K1141 NZ Coulomb sink under Phase 5k decomposed analysis but not sufficient under Boltz-2 full-likelihood scoring.

Body / ortho decomposition (n small — interpret carefully)

Body (n=12 → adamantyl 8, 1-indanyl 3, 4-F-biphenyl 1):

  • 4-F-biphenyl +0.035 (n=1) — unconvincing alone
  • adamantyl −0.007 (n=8) — neutral on average; head-group dominates within
  • 1-indanyl −0.019 (n=3) — slight negative

Ortho (n=12 across 6 substituents):

  • −CF3 +0.067 (n=1), −CN +0.037 (n=3), −Cl −0.008 (n=3), −F −0.041 (n=3), −H −0.053 (n=1), −Me −0.061 (n=1)
  • Pattern: electron-withdrawing ortho (CF3, CN) trends mut-prefer; small / electron-neutral ortho (H, Me, F) trends WT-prefer. −Cl is borderline.
  • n is too small per ortho group to confirm a rule; head-group signal is the cleaner finding.

Forward chemistry direction (data-grounded)

Tier-1 candidates (🟢 mut-prefer on Boltz-2, scaffold = aq3):

IDΔmech rationalenext gate
aq3__1-indanyl__phosphonate__-CF3+0.067−2 charge → stronger Coulomb sink; addresses DR4 mutagenicity (no Lossen); 1-indanyl saturates aromatic ring → DR4 photosensitivity OKPhase 5d ensemble re-dock + APBS — confirm K1141 NZ engagement under MD
aq3__adamantyl__CONHOMe__-Cl+0.066methyl-ester / hydroxamate prodrug (esterase activates locally → CONHOH at pocket); adamantyl already wet-lab-validated tailADMET re-check on CONHOMe (was in v5.2 library; verify flag-free); literature on cochlear vs plasma esterase differential
aq3__adamantyl__CONHOMe__-CN+0.065same prodrug logic; CN ortho electron-withdrawing — supports head-group activation chemistrysame gate as above
aq3__4-F-biphenyl__phosphonate__-CN+0.035second phosphonate hit, biphenyl tailPhase 5d re-dock; compare K1141 engagement vs 1-indanyl variant

Tier-2 — aq3__adamantyl__CONHOH__-Cl (current paper-claim lead): demote from #1 to reference / mech-anchor compound. Its role is to keep Phase 5k load-bearing (proves the K1141 +4.53 kT/e pocket attraction). It is not the optimal Boltz-2 picked binder. Paper §5 should explicitly distinguish “mechanism proof” (CONHO⁻ ↔ K1141 NZ attractor) from “lead optimization” (CONHOMe / phosphonate beats CONHOH on full-likelihood scoring).

Demoted — anything with COOH or CONHSO2Me heads on aq3-adamantyl body (5 ligands tested, 4 WT-prefer, 1 tie). DR4’s “design-around to acylsulfonamide” recommendation does not survive Boltz-2 evidence on this scaffold.

Caveats

  • n is small per group. 12 paired ligands across 5 head-group classes — strongest classes have n=4 (CONHSO2Me) or n=2 (everything else). Replicate top hits with 2–3 ortho variants each before any chemistry commit.
  • Boltz-2 ligand_iptm is a Boltz-internal confidence metric, not a binding-affinity prediction. Δ in ligand_iptm correlates with mut-prefer signal in Phase 7H NSAID benchmark but absolute thresholds are batch-relative.
  • Phase 5k decomposed-electrostatics proof remains load-bearing for mech 4 — that is a thermodynamic component (formal-anion preference, p=6.9e-12) and applies to any anionic ZBG. CONHOMe is neutral; if the v5.3 lead becomes CONHOMe, paper-§3 mech proof needs to either (a) keep CONHOH as the mech-proof reference compound and CONHOMe as the lead candidate (clean split), or (b) re-run Phase 5k-style APBS decomposition on CONHOMe geometry to verify a different (likely van-der-Waals + H-bond) mech-driver class.

Connections

Graph View